We develop a weather forecasting model ConvCastNet, which is based on convolutional neural networks. The training data are obtained from ERA5 reanalysis and are interpolated to an equiangular grid with a 3° spatial resolution. The forecast is carried out with autoregression of daily averaged variables, encompassing 10 atmospheric, sea and land variables and 3 static fields. A novel approach to the treatment of boundary conditions in convolutional neural networks is introduced, effectively mitigating challenges associated with the matrix representation of data. We conduct a comprehensive analysis of convolutional neural networks components and their training algorithm’s performance, thus optimising the forecast model. ConvCastNet achieves significant success in predicting geopotential at 500 hPa pressure level, with 7.8 days of useful forecast, defined by the lead time at which the anomaly
correlation coefficient is greater than 0.6. This result comes close to the results of the latest machine learning models and the deterministic high-resolution model (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF). ConvCastNet accurately predicts tracks of tropical cyclones for several days in advance and adjusts their intensity proportionally to ERA5 reanalysis. Nevertheless, due to the daily averaging of the data and a spatial resolution of 3°, the forecasts have limited utility, as predicting current and local weather characteristics holds greater significance in assessing the impact of such extreme events on people. An analysis of ConvCastNet forecast errors in the year 2022 demonstrates that their magnitude is greatest in highly baroclinic regions in high latitudes. Normalizing the errors with the atmospheric natural variability reveals that the error growth relative to the typical rate of local weather variability is greatest in the tropics.
|
